False-positive safety flag on routine product/coding conversation (Fable 5 -> forced Opus 4.8 switch)

Summary

During a normal software-development session (managing a web-app project, discussing task priorities with the assistant), a completely benign assistant message was flagged by Claude Fable 5's safeguards, and the session was force-switched to Opus 4.8 with an on-screen safety notice.

What was flagged

The flagged assistant message was a routine product/engineering status update:

• a priorities table (which tasks are done, which are pending)
• notes on a speech-to-text (STT) routing feature
• a discussion of tuning the length of generated hint text

There was no security-exploit content, no biology, no CBRN, no dual-use material of any kind. It was ordinary product management and coding work.

Why this is a false positive

Nothing in the message approaches any safety boundary. The in-product notice itself acknowledges this class of error:

"The safeguards are intentionally broad right now and may flag safe and routine coding, cybersecurity, or biology work. These measures let us bring you Mythos-level capabilities sooner, and we're working to refine them."

This was exactly that: safe, routine coding/product work incorrectly flagged.

Impact

• Interrupts a legitimate workflow with no indication of what triggered it.
• Forces an unwanted mid-task model switch (Fable 5 -> Opus 4.8).
• Reads to the user as an accusation while they are doing ordinary work, which is frustrating and erodes trust in the tool.

Requests

1. Tighten the false-positive rate for routine coding / product-management conversations.
2. When a message is flagged, surface (even coarsely) which part triggered it, so users can distinguish a genuine block from noise.
3. Provide a lightweight in-product "this was a false positive" feedback control tied to the specific flagged turn.

Environment

• Claude Code (desktop), Windows 11
• Model: Claude Fable 5 (auto-switched to Opus 4.8 on flag)